Industrial Robotic Perception

Few-/zero-shot anomaly detection and open-vocabulary grasp pose estimation for industrial robotic manipulation. Hyundai–NTU–A*STAR Corporate Lab · Nov 2025 – Present

Funding: Hyundai–NTU–A*STAR Corporate Lab · Nov 2025 – Present

Building the perception stack for reliable robotic part picking at HMGICS: open-world object detection and segmentation, 6D pose estimation, and grasp pose generation for cluttered industrial scenes, alongside foundation-model-based anomaly detection for unseen part categories.

Contributions:

  • Developed DriftAD, a visually-guided CLIP adaptation framework for few-shot industrial anomaly detection, achieving up to 1.5-point gains in 1-shot AUROC and PRO (ACM MM 2026).
  • Extended the work to zero-shot anomaly detection with TRACE, combining CLIP semantic evidence with DINOv3 structural features for cross-model anomaly detection without target-domain examples (submitted to AAAI 2027).
  • Built an open-vocabulary perception and grasping pipeline predicting detections, instance masks, and grasp poses for object categories unseen during training, validated across 77 categories of real industrial parts with an on-site robotic grasping demo.
  • Filed a patent for grasp pose estimation on wrapped parts in unstructured industrial environments.

Related publications: ACM MM 2026